The four levels of building AI systems

Build or buy is the wrong first question. Most companies have already bought. The real decision is which of four levels each AI tool belongs on.

Dušan StamenkovićFounding & Managing Partner · AI Practice LeadPublished 30 September 2026 · 11 min read

The idea in one minute

  1. 1Every AI tool sits on one of four levels of build: the subscription as bought, the configured subscription, the No-Code Track, and the Custom Build Track.
  2. 2Most of the value of the seats companies already pay for sits at Level 2, the configured subscription, and it is the level most skip.
  3. 3Choose the level for each tool, not once for the whole company. Who hears the answer, how many sources it depends on, and what a wrong answer costs decide it. A custom build is right when the goal requires it, or when it pays over the full lifetime cost.

01

Build or buy is the wrong first question

An elevator has four buttons. Most companies step in, see that the button for floor 1 is already lit, and reach straight for 4. Floor 1 is the AI subscription they already pay for. Floor 4 is a custom-built system. Floors 2 and 3, configuring that subscription and letting teams build their own tools, never get pressed.

Companies, and enterprises most of all, look at an AI investment as a choice between those two options. Nothing in between. The industry measures it the same way.

76%

of enterprise AI use cases were counted as bought in 2025. The survey sorts every use case into bought or built.

BoughtBuilt
Source: Menlo Ventures, annual survey of enterprise AI buyers, 2025

By the time a company asks me build or buy, it has usually already bought: seats for whole departments, used as an empty chat window. What leadership plans to build next is often something those seats could do once configured, or something a team should build itself.

The first question is not build or buy. It is which level each tool belongs on.

02

Four levels, one ladder

They differ in who does the work, what gets built, how many sources an answer depends on, who catches a wrong answer, and what kind of cost they carry.

1The subscription as boughtTime to build · Same dayWho does itEach personWhat gets builtNothingWho catches a wrong answerThe person using it, before they use the answerType of costSubscriptionsWho keeps the data from useThe vendor, under its termsMore on Level 1 ↓2The configured subscriptionTime to build · WeeksWho does itA central admin or platform teamWhat gets builtA company AI workspace: connections, permissions, onboarding, company knowledgeWho catches a wrong answerThe person using it, before they use the answerType of costSubscriptions, plus the team’s time to configure and maintain themWho keeps the data from usePartly you: your knowledge base, and the usage data you can exportMore on Level 2 ↓3No-Code TrackTime to build · Weeks per toolWho does itBusiness teams, with no-code builders or AI coding assistantsWhat gets builtSpecific tools for specific decisions, on one or two sourcesWho catches a wrong answerThe team member, before anyone acts on itType of costBuilder or coding-assistant subscriptions, model usage, metered calls into connected systems, and build timeWho keeps the data from usePartly you: the run logs and outputs of your own toolsMore on Level 3 ↓4Custom Build TrackTime to build · Months per systemWho does itEngineersWhat gets builtProduction systems that reconcile several sources, act on records and serve customersWho catches a wrong answerOften nobody before it takes effect, so it has to be tested to be rightType of costA project budget, then running and maintenanceWho keeps the data from useYou, completelyMore on Level 4 ↓

Times to build are for one team’s first version. At Level 2 the time goes into connecting sources, access, permissions and rules. Across an enterprise, rollout depends on scope.

The dividing line

Who catches a wrong answer? At the first three levels, a person reads the answer before it is used. At Level 4, the answer often reaches a customer or changes a record first. That, and how many sources an answer depends on, separate the levels more than anything else.

Each level also depends on the one below it. No-Code tools use the connections and permissions set up at Level 2. Custom systems are easiest to adopt when employees can reach them from the Level 2 workspace they already use. Skipping a level usually means building its parts later, inside a more expensive project.

Who keeps the data from use is the question most companies ask last. It decides what one system can pass on to the next. More in Which level keeps your data? and The model isn’t your moat. The loop is.

03

Level 1: the subscription as bought

Most companies are here. Microsoft alone reported more than 30 million paid Microsoft 365 Copilot seats in its July 2026 results. Enterprise seats across the main vendors list at roughly $20 to $30 per user per month on annual plans, as of Q3 2026.

The gap between access and use · US employees

Organization has integrated AI tools47%
Use AI a few times a week or more30%
Use AI daily15%
Source: Gallup, survey of 22,573 US employees, July 2026

At Level 1, the assistant is an empty chat window. It knows nothing about the company unless someone pastes it in, and its value is limited by what each person knows to ask.

Level 1 pays for itself per seat almost immediately. The company only sees that return when the saved time goes somewhere.

36min

a month is all a $30 seat must save, when an hour of that person’s time costs $50.

Source: Prosperaize calculation

42%

of frontline employees who use AI regularly save eight hours or more a week.

Source: BCG, 2026

66%

of them get limited or no guidance on what to do with that time.

Source: BCG, 2026

Gave AI’s impact on their productivity the top rating · US employees

All weekly users17%
With frequent use, a clear plan and a supportive manager50%

Almost three times as many give it the top rating once the use comes with a plan and a manager behind it.

Source: Gallup, July 2026

04

Level 2: the configured subscription

Level 2 is the same subscription, set up as a company AI workspace. It comes from four pieces of ongoing work:

  1. 01Connections to where company documents and records live.
  2. 02Permissions on who can reach what through the assistant.
  3. 03Onboarding that gives each role a short list of uses tied to its weekly work.
  4. 04Company knowledge kept current, with one owner per domain and one authoritative version of each document.

This is where most of the value of the seats sits.

Points added to the share of regular users reporting measurable impact

Clear direction on what to use AI for+25
Better tool access alone+5

Only about a third of employees felt they had been trained enough.

Source: BCG, AI at Work 2026, 17 markets

Two decisions at this level are hard to change later.

The permission modelThe assistant sees whatever each user can already see, so overshared files become visible through it.
The plan tierOn ChatGPT and Claude, limiting which roles reach which connected systems requires the Enterprise plan.

Level 2 needs an owner and a budget line, because none of it ends.

A common Level 2 use: enterprise AI search over company knowledge →

05

Level 3: the No-Code Track

On the No-Code Track, business teams build specific tools themselves, on data that already exists. They do it in one of two ways: with no-code builders such as Copilot Studio or n8n, or with AI coding assistants that let someone who knows the work write a small program. Examples from different functions:

SalesA one-page customer brief before every renewal call, with open tickets, usage and recent invoices.
FinanceA month-end checklist that flags missing entries from the ledger export.
HRAnswers to policy questions, with the policy section and its date.
Customer supportDraft replies built from the knowledge base and the ticket history.
Audit and qualityThe procedure that applies to a finding, with its revision and owner.

The tool costs are subscriptions for the builders or coding assistants, model usage, and metered calls into the systems a tool uses. On Copilot Studio’s pay-as-you-go rate, a prompt grounded in company data costs about 12 cents, and nothing for employees who already hold a Copilot seat.

The real cost is ownership. A tool written as code adds the obligations of software: someone who can review it, tests, and a place to run it that IT knows about.

43%

of breached organizations had security incidents involving AI tools employees use without approval, up from 20% a year earlier.

Year before
2026
Source: IBM, Cost of a Data Breach Report 2026
A common Level 3 use: workflow automation on data you already have →

06

Level 4: the Custom Build Track

At Level 4, engineers build, test and run production systems. In the right cases, it is the AI investment with the largest return.

Requirements

  • Customers hear the answer
  • The AI goes into a product you sell
  • The answer has to reconcile several systems that disagree
  • The system writes into your records
  • Your data is not allowed to go to a cloud assistant

Economics

  • Seats for large groups who use the assistant only occasionally
  • Metered usage at millions of prompts
  • An asset competitors cannot buy

Level 4 for privacy and scale

Sometimes the custom build is the assistant itself: a private assistant on your own cloud or on-premises, with the same connections, permissions and rules as Level 2. It is worth it when the data can’t leave your environment, when you need independence from one vendor, or when tens of thousands of seats cost more than running your own.

ChatGPT Enterprise, Claude and Copilot: when to configure them and when to run your own →

From my work before Prosperaize

On an e-procurement platform, the client asked for a recommender that matched suppliers to tenders. Before it could work, I had to build two systems nobody had planned: one for structured tender descriptions and one for supplier profiles. The feature was the recommender. Most of the work was the data it needed.

Custom investments start in two ways. Some grow out of a lower level, when a No-Code tool or a subscription agent outgrows it. Others start custom from day one, because the AI goes into a product, large-scale use is expected from launch, or the problem is one no subscription can reach.

Either way, the decision is made on the full lifetime cost, including maintenance. You either keep paying a vendor, build an internal team, or start with a vendor and plan the handover to your own team. If you want the capability in-house, write the handover into the vendor contract from the start.

That is why it is built only when the numbers hold, and tested before it is built.
A common Level 4 use: AI product development, inside what you sell →How we prove a build first →

07

Choose per tool, not per company

Choose the level for each tool, not once for the company. Most companies end up with tools on all four levels at the same time. Five questions decide it.

Who hears the answer?

Lower levelAn employee who checks it before actingLevel 4A customer, or a process that acts on it automatically

How many sources does it depend on?

Lower levelDocuments (Level 2), or one or two structured sources (Level 3)Level 4Several systems that disagree and have to be reconciled

What does a wrong answer cost?

Lower levelLittle, because a person catches itLevel 4A penalty, a safety or quality issue, a lost customer

Does it read or change records?

Lower levelReadsLevel 4Writes into a system of record

Can the data go to a cloud assistant?

Lower levelYesLevel 4No, because of network, residency or regulatory rules

Two more signals help between Levels 2 and 3. If the same role asks the same structured question many times a week, it is a candidate for a No-Code tool. If the answer depends on structured data rather than documents, it usually needs one.

The same use case can sit on any level. Usually one fits best, and the answers above decide which.

One use case, four levels

Enterprise AI search
Level 1Each person uploads files to a chat, one at a time.
Level 2UsuallyYour documents connected, with permissions, so answers cite the source.
Level 3A team-built assistant for one knowledge domain.
Level 4SometimesSearch across systems and formats a subscription can’t reach.
AI customer service
Level 1Agents ask a chat for help with a reply.
Level 2Sometimes · complex casesExpert assist on your knowledge base. A person sends every reply.
Level 3A team-built tool that drafts replies for one queue.
Level 4Usually · routine volumeRoutine requests resolved end to end, with a handoff to a person.
Workflow automation
Level 1A document pasted into a chat, the answer retyped.
Level 2Drafting and summarizing in the tools you have, set up for the team.
Level 3Usually · one teamA team-built tool reads, sorts and routes on one or two sources.
Level 4Usually · enterpriseEngineers automate a workflow across several systems.

Filled: where it usually sits. Dashed: where it sometimes does. Each use case links to its page.

08

Does it pay at that level?

One test, every level

The decision a tool improves, times how often that decision is made, has to be worth more than what the tool costs plus what its wrong answers cost.

What the tool costs depends on the level:

Levels 1 and 2The subscriptions, which most companies already pay for.
Level 3Subscriptions for the tools it is built with, such as Copilot Studio, n8n, Lovable or an AI coding assistant; model usage, often priced per token; metered calls into the systems it uses, such as Jira or Confluence through their MCP servers; and the time spent building it.
Level 4The build, the data work, running costs and maintenance, over three years.

Say an account manager’s renewal brief saves one lost renewal a year. The No-Code tool behind it costs a few subscriptions and usage fees, plus a few days of someone’s time to build. It passes easily.

How much analysis to do before investing, and how much testing before relying on the result, also depends on the level:

1The subscription as boughtAnalysis before you investChoose the platform your company already works in. No business case needed.Testing before you rely on itLet a few people use it for a few weeks and check whether they keep using it.
2The configured subscriptionAnalysis before you investChoose the first roles and their uses, the documents those uses need, and which sites need permission cleanup.Testing before you rely on itStart with two or three roles for a month. Collect wrong answers and send each to the owner of that domain.
3No-Code TrackAnalysis before you investA one-page estimate: the decision, how often it is made, what it is worth, the tool costs, the build time and the owner.Testing before you rely on itThe builder tests it on real past cases. If it is code written with an AI coding assistant, someone who can read code reviews it. A person checks every answer until the tool has proven itself, then reviews it on a schedule.
4Custom Build TrackAnalysis before you investA full business case: three-year costs, the cost of wrong answers against today’s, and who maintains it. During the proof of concept, the unit economics of each answer.Testing before you rely on itTest the riskiest assumptions on a prototype, with a decision to stop or continue, before the full build. Before launch, test against a set of real questions with known answers.

09

Every level needs an owner

Each tool needs a named owner and a place on one company-wide list: what it does, which decision it supports, which data it uses, who uses it, and when it was last checked. The list is also where you notice a tool that has started to need a higher level.

Before a tool is built, get three people into one room:

01The person who knows the day-to-day operation
02The person who knows the data
03Someone who knows and owns AI

Agree on what gets built, what data it needs, how complex it is, which level it belongs on, and how to get the most performance for the cost and the risk.

10

Is this for you?

This is for leaders deciding where to put the next AI budget: which tools to configure, which to let teams build, and which to fund as a custom system.

If the AI budget exists because the board asked for one, start by listing the decisions you want to improve. The levels only help once there is a decision to place.

Place your own tools.

Schedule a call with a partnerFree until 11 December